license: mit base_model: orcarouter/GLM-5.3-Flash-Uncensored-FP8 base_model_relation: quantized pipeline_tag: text-generation tags:

• gguf
• quantized
• glm
• iq2
• iq3
• uncensored
• project-maya language:
• en
• ja

GLM-5.3-Flash-Uncensored — Optimized IQ2/IQ3

A calibrated, mixed IQ2/IQ3 text-model conversion of orcarouter/GLM-5.3-Flash-Uncensored-FP8, built and tested on an NVIDIA RTX PRO 6000 Blackwell 96 GB, an Intel i9-12900KS, 128 GB RAM, and NVMe storage.

The two GGUF shards occupy 89,879,704,352 bytes (83.707 GiB). With the Maya pack, shared dense weights, and tokenizer, the runtime model directory occupies approximately 91.866 GB (85.557 GiB). Both GGUF shards and the pack/ directory must stay together.

The upstream model name includes Uncensored; this conversion keeps that identity. This is the completed compact model deployed for planning, coding, and Japanese creative text in Mang-AI.

Quantization

• Routed experts mainly use IQ2_XXS, with selected important down projections in IQ3_XXS.
• Retains the full 45-layer text trunk and all 288 experts in each of 42 routed layers. Experts are not pruned or skipped.
• Router, mHC, KDA, indexer, and absorbed MLA weights are stored in F32. Other shared weights mainly use Q6_K/Q8_0.
• Converts the pinned FP8 source directly, without materializing a full BF16 intermediate checkpoint. Calibration and GPTQ error feedback were used.
• MTP and the vision tower are excluded from this single-GPU text-only build.

This is a low-bit compact conversion, not a predominantly 4-bit model. It has not been shown to match the original FP8 model's quality.

Source revision: 0480f690f80e8f4c4ee63edd10dcbea8b7d9e31b.

Files and runtime compatibility

GLM53-Compact-00001-of-00002.gguf
GLM53-Compact-00002-of-00002.gguf
pack/
  dense.bin
  index.txt
  native_experts.txt
  expert_prior.txt
  expert_counts.txt
  tokenizer/

pack/ is a Project Maya native-expert pack, using GGUF shards as the backing storage for experts. It does not duplicate the full expert weights. Keep the shard filenames unchanged: the native-expert index references them by name. Runtime telemetry and usage sidecars are deliberately omitted.

The measured configuration uses a locally patched Project Maya runtime based on commit 6083122531be0b9c5d22b663e355916840b0c927, with GLM native-expert v3 shard support and precision-preserving shared-weight handling. The three production patches are included in runtime/patches/, with the upstream runtime license in runtime/MAYA-LICENSE. Apply these patches to the pinned Maya checkout before building its CUDA runtime; the maya-f32-runtime-views.patch adds the precision-preserving shared-weight compute views. Follow Maya's build instructions and use the recorded llama.cpp revision from upstream.lock.json. Building or installing that compatible runtime is a separate step. Compatibility with stock llama.cpp, LM Studio, Ollama, or an unpatched Maya build has not been established. Loading the two files alone in an arbitrary GGUF runner is not the tested deployment.

Download the repository with hf download darask0/GLM-5.3-Flash-Uncensored-IQ2-IQ3-Optimized --local-dir ./glm53-uncensored. SHA256SUMS and model-manifest.json contain file checksums. quantization.json records the recipe. runtime-template.json records the tested settings with placeholder executable/model paths; replace the paths for your runtime before use. Settings are sized for the 96 GB GPU above and are not a universal hardware preset.

Tested settings: 32,768 context, FP16 latent KV, FP32 KDA state, 2 GiB prefill lending, prestage disabled, KDA graph disabled, all experts retained. File size does not include KV, execution buffers, temporary compute copies, or the display's VRAM usage. The local launch checks required at least 90.07 GiB free VRAM and 17.45 GiB free RAM. Observed process PSS after deployment and media checks was 11.489 GiB; this is a snapshot, not peak memory.

Measured validation

Median of three runs at each input length, excluding warmup; temperature 0, reasoning disabled, 256 output tokens, cache_n=0. API wall time includes input processing and generation.

These figures describe this PC and test configuration; they are not a promise for other hardware or the usual UI's reasoning-high setting. A 42.47-second load observation was not a controlled cold-start measurement.

• API smoke: 4/4 passed.
• Isolated coding checks: 10 tasks, 58 cases passed.
• 16K and 30K inputs: fixed identifiers were retrieved from three positions. This tests retrieval, not general long-document comprehension.
• Distribution comparison: 8 sequences, 7,672 scored tokens; NLL 2.10669 versus reference 2.14849, top-1 agreement 77.36%, top-64-plus-tail KL 0.77662. The reference used FP8-derived BF16 hidden streams and FP16 output projection. Individual sequences worsened; the average NLL does not establish FP8-equivalent or superior quality.
• Real coding-agent/tool and structured Japanese four-panel output checks passed in the local deployment.
• GLM unload/reload around image generation, video generation, and a one-step LoRA backend test was verified in Mang-AI. These are integration checks, not image/video/LoRA quality evaluations.

No full standard benchmark suite or 4-bit-model comparison was completed. This text-only release does not perform image recognition.

日本語

このPC用に作成・検証したGLM 5.3 Flash UncensoredのIQ2/IQ3混合版です。GGUF2本で83.707 GiB、pack込み約85.557 GiB。司令塔・コーディング・日本語創作テキストに同じモデルを使用しています。全expertを保持し、重要な共有重みを保護しています。元FP8と同等品質、汎用GGUFランナーでの互換性、別ハードウェアでの速度は未確認です。

License and attribution

The source model is distributed under the MIT license; its original copyright notice is included in LICENSE. The base weights are credited to Z.AI and the upstream Uncensored FP8 repository to orcarouter. Runtime/source references and revisions are recorded in upstream.lock.json.